Approaches

Baselines

The ladder needs a floor that is obviously weak and a second rung that is obviously stronger, so that the arena's first record checks the machinery rather than any idea about Stratego.

The players

random chooses a uniformly random setup and a uniformly random legal move. greedy scores every legal move one ply ahead: the expected material of an attack under the counting belief (a known weaker piece is taken, a hidden piece is weighed by the probability the attacker survives), a small bonus for advancing, a penalty for stopping next to a known stronger piece, and a probe bonus for Scouts attacking unknowns. Both use random setups. Greedy has no memory and no search; it is the strongest player that costs nothing.

What the first record has to show

Random against greedy on 200 paired games under the training rules. The expectation is lopsided, and the point of the record is the rest of the table: the endings histogram (how many games greedy ends by capturing the Flag against by leaving random with no moves), the mean length, and that the paired-seed contrast is symmetric when random plays itself. A random self-match that is not near 50% with a covering interval would mean the seeding or the seat pairing is wrong.

Decision rule

Greedy is the reference for the next rung if its effective win rate against random exceeds 90% with the 95% Wilson interval above 80% on all 200 games. If it does not, the greedy heuristic is fixed before anything above it is measured, because a search that beats a broken greedy has proved nothing.

Evidence

Run on September 18, 2026: greedy won 180, drew 1 and lost 19 of 200 paired games against random, 90.2% effective with a 95% Wilson interval of 85.4% to 93.6%; the random self-match scored 47.0% (40.2% to 53.9%) for the first seat set. The decision rule is met and greedy is the reference. See the report and the run record.

What could go wrong

Random play under the 100-move rule draws often, so the effective win rate of greedy may be dragged toward a half by draws it fails to convert rather than by losses. The record separates draws from losses for that reason, and a greedy that draws more than a quarter of its games gets a conversion term (prefer attacks when the battleless counter is high) before being used as a reference.